how many predict objets per cell?

[PythonImageDeveloper]

Hi ,
I have some question and i don't understand well.
in the yolo1 we consider 2 boxes per cell and each center objects located that certain cell , thus boxes of that cell assigned to those objects , my question is : each cell assigned to only one object or assigned to those objects located on that cell ?
why yolo1 can't detect small objects ? because use 7x7 ? this problem addressed in the v2 / v3? how ? only because grid cell 13x13 ? or because of B=6 instead of B=2?
and final question is in the v1 we predict bbox coordinates or offsets?

[experiencor]

1. Object => center of object => cell => best IOU => box. Assign to all objects located on that cells.
2. Small objects do matter. What matters is that one cell is only able to predict a limited number of objects. If too many small objects fall on the same cell, many of them will not be detected.
3. To address this problem, increase the number of boxes and number of cells in each cell (YOLOv2, 13x13 grid, 5 boxes in each cell) or detect at multiple resolutions (YOLOv3, 13x13, 26x26, 52x52 grids, 3 boxes in each cell).
4. I'm not sure about v1 but v2 and v3 predict offsets.

[PythonImageDeveloper]

Thanks.
if center of two objects fall on the same cell , and we have 5 boxes (box1,box2,...,box5) per cell and assume the box1 assigned to the object1 , in this case the box1 can be assign to the object2 again? or the rest of boxes (box2,box3,...box5) just only can be assign to the object2?
Small objects do matter. What matters is that one cell is only able to predict a limited number of objects. If too many small objects fall on the same cell, many of them will not be detected.
why did limit the number of objects that can predict by same cell ? what is the problem if we set a large number of objects than can predict by same cell? How many number of boxes per cell did you have ?

Q2) Object => center of object => cell => best IOU => box
in this matching strategy only one box assigned to object fall into same cell or each boxes that have IOU above certain threshold can assigned to that object . that is mean not just only one boxes assigned to object?

[experiencor]

Objects are assigned randomly to the boxes. Overriding does happen at times.

Need to limit the number of boxes so as to limit the size of the network. Increasing the number of boxes of your objects are too small. It's perfectly fine. The standard model has 5 objects.

The box that has the max IOU with an object is responsible for predicting that object.